A Device for Computer Vision Analysis of Fungal Features Outperforms Quantitative Manual Microscopy by Experts in Discerning a Host Resistance Locus

Abstract

Accurate, quantitative phenotyping aids in the discovery of quantitative trait loci, particularly those with minor effects. Previously, we optimized replicated precision phenotyping of mapping families after inoculation of leaf discs with the grapevine powdery mildew pathogen ( Erysiphe necator). Pathogen colonies were stained, and hyphal density was estimated using hyphal transects. This approach outperformed field evaluations and other controlled phenotyping methods but required one or two person-months of microscopy per experiment to evaluate resistance across 300 host genotypes. More recently, we combined advanced macrophotography, robotic sample positioning, and convolutional neural networks to produce a high-throughput phenotyping device, which was modified and commercialized as “Blackbird.” Here, that device was tested for nondestructive image collection and computer vision quantification of foliar grapevine powdery mildew. Blackbird outpaced manual microscopy up to 60-fold and nondestructively generated time-series segregating phenotypes from 2 to 9 days postinoculation (dpi). Paired analysis of these phenotypes with RNase H2-amplicon sequencing haplotype markers targeting the Vitis core genome detected REN13 on chromosome 8. Genetic analysis of Blackbird convolutional neural network data explained a greater proportion of the phenotypic variance via hyphae at 4 dpi (24.5%) and conidia at 9 dpi (24.0%) than manual microscopy at 8 dpi (15.8%). As a moderate-effect resistance locus in the widely planted resistant variety ‘Norton’, which already produces commercial wine quality, REN13 could significantly delay epidemics and could be useful in grape breeding programs to increase the durability of stronger resistance loci (e.g., RUN1, REN4, or REN12) in resistance gene stacks while maintaining fruit quality.